I. Introduction
The UKF was proposed by Julier and Uhlmann, which uses a deterministic sampling approach to avoid sub-optimal performance and divergence of the extended Kalman filter (EKF) [1]. Then, Eric and Rudolph extended the use of the UKF to nonlinear estimated problems [2]. From then on, the UKF has been widely used in many applications, such as target tracking, vehicle navigation, neural networks training, and so on. With the development of sensor networks, centralized and distributed UKF s [3] [4] are designed for the networked environment. Although the former method has high precision, the robustness of the system is low. Once the local sensor fails, the entire network will be paralyzed. Compared with that, the distributed one has a better performance in stability and robustness to failure, and structural flexibility.